[P0] Helpfulness-driven rerank weight tuning v1

[justinbuzzni]

Goal

• Improve retrieval quality over time using user/session helpfulness signals.

Scope

• Collect helpful/neutral/unhelpful feedback signals.
• Periodically tune lexical/semantic/recency weights.
• Add rollback guard when quality drops.

Acceptance Criteria

• Top1 +5%p OR Top3 +8%p improvement on replay benchmark.
• Automatic rollback path exists.

Checklist

• Define feedback schema
• Implement tuner job (v1)
• Add rollback guardrail
• Produce before/after quality report

[justinbuzzni]

Progress update:

Implemented helpfulness-driven adaptive rerank weights (v1):

• Added optional adaptive rerank in MemoryService.retrieveMemories(..., { adaptiveRerank: true })
• Weights are derived from helpfulness stats (avgScore, helpful/unhelpful balance)
• Retriever now supports rerankWeights overrides (semantic/lexical/recency)

Validation:

• npx vitest run tests/retriever-fallback-chain.test.ts tests/retriever-strategy-scope.test.ts tests/retriever.memu-adoption.test.ts ✅ (7/7)
• npm run build ✅

Commit:

• 7de0296